#Tips om free working models 🙏

1 messages · Page 1 of 1 (latest)

unique lintel
#

I’ve been trying to get openclaw to work for like 4 days now. I can run it on paid apis, but it drained my credits so quick that as a broke student I quickly realised this was not doable. I followed some guides online to keep the context low and clear every time it passed x tokens, even added configs and instructions for this, but with like a couple config messages once the bot was running, it drained 5€ of tokens with haiku and sonnet. Before moving to the free ones, I also tried Kimi k2.5 through openrouter, but same issue, a couple small short messages, really few actually, and it spent 2€

After this I decided to try and have it run on free models with openrouter, but so far all the models I tried did not work. I tried Pony and Gpt-oss on open router, both did not recognised the model ids. I tried Llama 70b with groq api on the free tier and the Same thing happened. I then tried Kimi k2.5 through nvidia api, supposedly, also not recognised.

It’s been very frustrating and almost giving up honestly. Anyone could help me out and give me tips for working free models on open router or another provider?

Also, some tips on how to actually use paid models without burning tokens sooo quickly? Bc from what I’ve been reading, people can spend around 20€ per month even with paid ones, that is doable for me, but from my experience it could easily be 20€ per day 😂 at least the way mine is configured

#

Why is it dismissed?

timid ploverBOT
# unique lintel I’ve been trying to get openclaw to work for like 4 days now. I can run it on pa...

This usually ends up being two separate problems: (1) model names (IDs / allowlist), and (2) token burn from context + tool schemas.

1) “Model id not recognized” / “unknown model” / “not allowed”

OpenClaw expects models as provider/model (all lowercased). A bunch of “short names” people post in guides (like “pony”, “gpt-oss”, etc.) are not real IDs unless you define them as aliases.

Do this instead of guessing:

  • In chat: /model list (shows what your current agent is actually allowed to pick)
  • On CLI:
openclaw models list --all
openclaw models list --all --provider openrouter
openclaw models scan --no-probe # OpenRouter free-catalog candidates (no live testing)

Two common gotchas from the docs:

  • If the model id itself contains / (OpenRouter-style), you must include the provider prefix when selecting (so OpenClaw can parse it).
#
  • If you set agents.defaults.models (an allowlist), anything not in there will fail with “Model … is not allowed” and it can feel like “it didn’t reply”.

Docs: https://docs.openclaw.ai/cli/models and https://docs.openclaw.ai/providers/openrouter

2) Why “a couple short messages” can still cost €€

Even if your messages are tiny, each run can include:

  • the system prompt
  • injected workspace files (AGENTS/SOUL/TOOLS/etc)
  • and (big one) tool schemas (browser/exec/canvas/etc)

So you can burn thousands of tokens per message before you’ve “said” anything.

Quick way to see what’s happening:

  • /usage tokens (adds a tokens footer)
  • /context detail (shows what’s dominating the prompt)

Docs: https://docs.openclaw.ai/concepts/context and https://docs.openclaw.ai/token-use

#

Fast cost reducers (big impact)

  1. Restrict tools (tool schemas are expensive)

  1. Disable heartbeat while debugging
    Heartbeat is a full model turn on a timer.
    In config: agents.defaults.heartbeat.every: "0m" (disable) or make it much longer.

  2. Be careful with “compact every X tokens” configs
    Compaction is another LLM call. If it triggers too often, it can become the thing that eats your budget.

If you paste 3 outputs, I can tell you exactly what’s happening

(These are safe-ish to share; --all redacts tokens.)

openclaw --version